6.2.1 Conceptual background and summary of the present findings
An examination of the largely separate literatures on neural oscillations in the alpha band and neuronal variability across trials reveals many parallels between the two measures of brain activity. Neural variability substantially decreases with the onset of a visual stimulus (Churchland et al., 2010), as does the amplitude of posterior alpha oscillations, a phenomenon known as event-related desynchronization (ERD) (Pfurtscheller and Lopes da Silva, 1999). The degree to which neural variability in human EEG recordings is quenched by visual stimulation is a stable individual characteristic that determines individual perceptual thresholds (Arazi et al., 2017b, 2017a), while individual perceptual performance is also predicted by the individual level of parieto-occipital alpha activity prior to stimulus onset (Hanslmayr et al., 2007). The detection of visual stimuli presented at the perceptual threshold is generally associated with reduced trial-to-trial variability (Ress et al., 2000; Schurger et al., 2015) as well as reduced alpha amplitudes (Ergenoglu et al., 2004; Hanslmayr et al., 2007; Dijk et al., 2008;
Benwell et al., 2017), and both neural variability across trials and alpha amplitudes are reduced for stimuli that are attended compared to unattended stimuli (Sauseng et al., 2005b; Kelly et al., 2006;
Thut, 2006; Mitchell et al., 2007, 2009; Cohen and Maunsell, 2009; Mo et al., 2011; Arazi et al., 2019).
Based on these striking similarities between the functional roles of neural variability and alpha oscillations in visual perception it stands to reason that the amount of trial-to-trial variability typically measured in neuronal spiking activity could at least partially be explained by the strength of low-frequency cortical oscillations.
In the dissertation presented here we showed that neural spiking activity in the visual thalamus, more specifically the lateral geniculate nucleus (LGN) and the dorsal and ventral pulvinar, does not exhibit reductions in variability upon stimulus onset to the same degree as in visual cortex area V4 examined in the same animals (Chapter II). Moreover, we found ongoing thalamic activity prior to stimulus onset to be considerably less variable than cortical activity (Chapter II), suggesting a dependence of the decrease on the amount of variability inherent in the spontaneous neural activity preceding visual stimulation. Investigating the effects of perceptual suppression on neural variability as well as on the amplitude of the simultaneously recorded local field potential (LFP) in macaque visual area V4, we
121
found a significant reduction in neural variability around the time of perceptual suppression that coincided with a significant reduction in the amplitude of LFP low-frequency oscillations around the alpha range (Chapter III). A comparison of alpha amplitude and spiking variability in the examined time window showed a positive site-by-site correlation between the two measures (Chapter III). In the human EEG, we found the individual degree of variability quenching to be closely correlated with the degree of individual reductions of alpha amplitude with stimulus onset (Chapter IV). Interestingly, both stimulus-induced decreases were also dependent on the amplitude of alpha band oscillations prior to stimulus onset and examining time courses of the raw variance across trials and alpha amplitude of individual subjects we noted that both measures closely co-varied (Chapter IV). The results presented here thus suggest two main conclusions which will be discussed in more detail in the following sections: 1) The amount of neural variability across trials largely reflects the amplitude of low-frequency oscillations, and 2) Their respective stimulus-induced decreases are dependent on the spontaneous activity dynamics preceding them.
6.2.2 On the relationship between alpha activity and neural variability
Neural variability is thought to primarily arise from ongoing fluctuations in cortical excitability that are highly correlated between large populations of neurons (Arieli et al., 1996; Goris et al., 2014; Schölvinck et al., 2015) (Arieli et al., 1996; Goris et al., 2014; Schölvinck et al., 2015). The magnitude of neural variability thus depends on cortical state, which likely exhibits a continuum ranging from highly synchronized states that are characterized by strong low-frequency oscillations, to more desynchronized states during which such activity fluctuations are largely absent (Harris and Thiele, 2011; Schölvinck et al., 2015). Neural oscillations in the alpha band specifically have been intensively studied and shown to have an inverse relationship to cortical excitability (Lange et al., 2013; Iemi et al., 2017; Samaha et al., 2017). Despite these functional parallels between alpha activity and neural variability, which further both vary with visual perception and attentional states in a similar manner, the two neural signatures have largely been investigated and reported independently. One recent study finally conclusively linked trial-to-trial variability and low- frequency oscillations: Daniel and colleagues could show that the two measures of brain activity are tightly coupled and that on the individual level the largest portion of inter subject differences was explained by individual power in the alpha and beta frequency range (Daniel et al., 2019). When the data were filtered with a band-stop procedure removing
122
the influence of a specific frequency band on the data, variability changes with stimulus onset were unaffected by the removal of the delta and gamma frequency bands, while neural variability increased relative to prestimulus baseline in the absence of alpha/beta frequency oscillations (Daniel et al., 2019).
These results are largely consistent with our examination of frequency contributions to variability quenching in the human EEG (Chapter IV, Supplementary information S3). Daniel and colleagues considered a frequency range spanning 5 – 25 Hz as the alpha/beta band since they did not observe any qualitative differences when separating alpha and beta frequencies (Daniel et al., 2019). This range partially includes theta (4 - 7 Hz), alpha (8 -12 Hz) and beta (13 - 30 Hz) frequencies, which we analyzed separately. We similarly found the impact of removing theta and gamma oscillations on variability quenching to be minimal, while considerably less variability quenching occurred without the alpha range (Chapter IV, Supplementary information S3). Band-stop filtering in the beta range did not affect variability quenching in our case (Chapter IV, Supplementary information S3). However, unlike Daniel et al. we also examined the absolute variance rather than the relative change in variance from prestimulus baseline and found that removing the influence of beta oscillations reduced both prestimulus variance as well as poststimulus variance to the same degree (Chapter IV, Supplementary information S3, Figure 2A). Prestimulus variance was most profoundly reduced following band-stop filtering in the alpha (8 – 12 Hz) range, resulting in a minimal stimulus-induced decrease to the same level of poststimulus variance observed for the removal of beta oscillations (Chapter IV, Supplementary information S3, Figure S2A). We also observed qualitatively similar changes in trial-to-trial spiking variability as well as simultaneously recorded theta, alpha and beta oscillations of the monkey LFP related to perceptual suppression (Chapter III). It is thus likely that while alpha oscillations, which are characteristically higher in amplitude than those in other frequency bands (Chapter IV, Supplementary information S3, Figure S2B), constitute the largest contribution to neural variability, this relationship is not exclusive to the classical alpha band. Concordantly, a recent study examining spiking activity related to movement execution in behaving macaques found reductions in trial-to-trial spiking variability to correlate with reductions in LFP beta power (Riehle et al., 2018).
The decline of trial-to-trial variability with stimulus onset could in theory also be explained by increased phase-locking of low-frequency oscillations rather than by a decrease in their amplitude, which we did not control for in our studies. Daniel et al. were however able to conclusively rule out this alternative explanation as they found inter-trial phase coherence to be uncorrelated with neural variability
123
quenching, suggesting that reductions in neural variability are in fact caused by reductions in oscillatory power (Daniel et al., 2019). Moreover, they found the spatial topographies of neural variability quenching and the ERD in the alpha/beta range to closely match and to be centered around the parieto-occipital areas we considered for our analyses (Daniel et al., 2019) (Chapter IV), further confirming the direct relationship between the two measures.
6.2.3 Are stimulus-induced changes determined by ongoing activity?
In the present dissertation we observed a strong correlation between individual prestimulus alpha amplitude and the degree to which neural variability across trials as well as posterior alpha amplitude declined with RDM stimulus onset, suggesting that the previously examined ERD in the alpha band as well as neural variability quenching may depend on the level of prestimulus alpha activity (Chapter IV).
The hypothesis is further supported by the finding that the removal of low-frequency oscillations in the alpha and to some degree in the beta range resulted in considerable reductions in prestimulus variance while poststimulus variance was only minimally affected (Chapter IV, Supplementary information S3, Figure 2A). Stimulus-induced decreases of neural variability have consistently been observed across various measures of brain activity including the membrane potential of cats (Churchland et al., 2010), spiking activity of non-human primates (Churchland et al., 2006, 2010; Oram, 2011; Poland et al., 2019) (Chapter II), human EEG / MEG recordings (Arazi et al., 2017b, 2017a, 2019) (Chapter IV) and the human fMRI (Broday-Dvir et al., 2018). The phenomenon of neural variability quenching is generally thought of as a network effect that arises from large neuronal populations independent of their responsiveness and stimulus specificity (Churchland et al., 2010; Rajan et al., 2010; Deco and Hugues, 2012; Goris et al., 2014) and improves the reliability of sensory processing (Zohary et al., 1994; Shadlen et al., 1996; Parker and Newsome, 1998). Here we showed that neural variability quenching does not occur to the same degree in the LGN and higher-order thalamus as in simultaneously recorded visual cortex V4, while spontaneous thalamic spiking activity prior to stimulus onset was already considerably more reliable than cortical activity (Poland et al., 2019) (Chapter II). This finding further hints at a possible dependency of ERD and variability quenching on the magnitude of ongoing activity fluctuations preceding them.
124
Previous studies have highlighted the role of ongoing activity in sensory processing, suggesting that evoked activity constitutes a linear summation of a deterministic stimulus response onto ongoing activity fluctuations prior to stimulus onset, which explain the large degree of variability in neural responses and its dependence on cortical state (Arieli et al., 1996; Schölvinck et al., 2015). Moreover, selective attention is observed to coincide with a desynchronization, which is a decrease in the amplitude of low-frequency oscillations, occurring in cortical regions that process the attended stimulus (Harris and Thiele, 2011). Similar desynchronizations of low frequency oscillations occur following external stimulus input and largely explain the phenomenon of variability quenching (Daniel et al., 2019). It could thus be argued that stimulus-induced changes in cortical state are dependent on the degree to which anticipatory attention is assigned to a given stimulus, which is conceptually supported by previous observations of larger magnitudes of variability quenching in response to attended compared to unattended stimuli (Arazi et al., 2019). A previous EEG study found only the relative variance, which is the amount by which variability decreases with stimulus onset, to differentiate between perceptual outcomes, while the absolute variance post stimulus was not perceptually relevant (Arazi et al., 2017a).
While the interpretation of negative results has to be approached with caution, this surprising finding could in theory be attributed to significant differences in prestimulus alpha amplitude brought about by anticipatory attention, which were not specifically examined within the same study and previously shown to predict subsequent stimulus perception (Chapter IV; Chapter V). Many previous studies have primarily considered relative changes with stimulus onset (Churchland et al., 2010; Arazi et al., 2017b;
Daniel et al., 2019), making it difficult to determine whether the improved perceptual performance associated with stronger variability quenching and ERD is mechanistically determined by neural activity dynamics prior to or following stimulus onset, which may further be strongly correlated. In order to impact sensory perception on a given trial, neuronal spiking variability would have to be apparent not only across trials but also in the temporal pattern present during stimulus processing, and stronger variability quenching has previously been associated with increased pattern stability within-trial (Schurger et al., 2015). Our findings comparing trial-to-trial variability between thalamus and cortex suggest that the magnitude by which variability quenches with stimulus onset is restricted by the degree to which the reliability of spiking responses can be further improved, resulting in reduced variability quenching in the thalamus wherein we found Fano factors of ongoing activity to already be close to 1 and thus close to the amount of variability expected to be inherent in neuronal spiking without the
125
additional impact of correlated activity fluctuations (Poland et al., 2019) (Chapter II). In a combined EEG-fMRI study, Becker and colleagues were able to show that the variability in evoked responses in the human fMRI is largely explained by ongoing alpha activity, which further strongly supports our hypothesis that the power of low-frequency oscillations preceding stimulus onset determines the reliability of subsequent stimulus processing (Becker et al., 2011). Increased cortical excitability associated with reductions in low frequency power as well as reduced variability following visual stimulation are conceptually likely to considerably improve stimulus processing, and the correlation between prestimulus alpha amplitude and ERD in the alpha band as well as neural variability quenching we observed (Chapter IV) hint at such dynamics underlying a common neural mechanism of anticipatory selective attention.